Net Revenue Retention for AI Startups: Formula and Pitfalls

How to calculate net revenue retention, how it differs from gross retention, why usage pricing distorts it for AI startups, and how to report it.

Sergei Mochtchenkov, CFA ·

New logos make the growth chart look good. Net revenue retention tells an investor whether those logos are worth anything a year later. For an AI startup raising a seed extension or a Series A, it is often the metric that decides whether the growth story holds up in diligence.

I'm Sergei Mochtchenkov, a CFA Charterholder and fractional CFO for AI-native startups. This guide covers how to calculate net revenue retention (NRR) correctly, how it differs from gross retention, why usage-based pricing makes it harder to measure for AI companies, and how to present it so it survives an investor's own recalculation.

What net revenue retention measures

Net revenue retention, also called net dollar retention, answers one question: of the recurring revenue you had from a group of customers at the start of a period, how much do you have from that same group at the end?

The standard annual formula is:

NRR = (Starting ARR + Expansion − Contraction − Churn) / Starting ARR

Each part is measured only for the customers who were paying at the start of the period:

  • Starting ARR is the annual recurring revenue from that cohort on day one.
  • Expansion includes upsells, additional seats, higher usage tiers, new modules and price increases from the same customers.
  • Contraction covers downgrades and reduced usage from customers who stayed.
  • Churn is the ARR lost from customers who left entirely.

Revenue from customers who signed during the period is excluded from both the numerator and the denominator. That is the whole point: NRR isolates what your existing base does without any new sales.

An NRR above 100% means the existing customers are spending more over time, so the company grows even before adding a single new logo. Below 100% means new sales first have to fill a hole before they add growth.

A worked example

Here is an illustrative cohort, measured over twelve months:

  • Starting ARR from customers active on January 1: 2.0M
  • Expansion from those customers during the year: 0.5M
  • Contraction from customers who downgraded: 0.15M
  • ARR lost from customers who cancelled: 0.25M
  • Ending ARR from the same cohort: 2.1M

NRR = 2.1M / 2.0M = 105%

Gross revenue retention for the same cohort excludes expansion: (2.0M − 0.15M − 0.25M) / 2.0M = 80%.

Both numbers matter. An NRR of 105% looks healthy, but an 80% gross retention means one in five dollars is leaking out each year and expansion is covering it. If expansion slows, the business feels the churn immediately.

Gross retention vs net retention

Bessemer's Scaling to $100 Million report draws the distinction clearly. Gross retention "nets out the revenue from customers who turned off or downgraded your service, but does not account for any expansion," while net retention accounts "for the upsells and expansion revenue" earned over the period.

The report also warns that "it is difficult to build a successful cloud business without having high gross revenue retention," and that "every percentage taken out of your retention is taken out of your growth rate." Its cloud portfolio data, published in 2021, showed the middle half of companies with net retention above 100% and noted that SMB segments tend to run lower than enterprise. Use it as context rather than a target, because AI-native companies have different pricing and churn patterns from the cloud companies in that dataset. My AI startup benchmarks page tracks current expectations for AI companies specifically.

Investors usually ask for both numbers together. NRR shows upside from the base; gross retention shows how durable that base is.

Why NRR is harder for AI startups

AI products bring three measurement problems that a seat-based SaaS company sees less often.

Usage revenue moves around. Many AI products charge per token, per task, per outcome or seats plus usage. A customer can double their spend in one month during a pilot and halve it the next. If you annualize a single strong month as ARR, both the starting base and expansion become unreliable. Decide what counts as recurring: committed contract value, a trailing average of usage, or a minimum commitment. Write the definition down and use it every period.

Pilots inflate the cohort. Paid pilots and proofs of concept often convert at a different rate from full contracts. Including them in starting ARR and then losing them makes churn look worse; excluding them and counting the conversion as expansion makes NRR look better. Either approach can be defended, but investors will want to know which one you chose and why.

Price changes can dominate. Model costs fall, and many AI companies change pricing more than once a year. A price increase shows up as expansion, while a price cut to stay competitive shows up as contraction even if usage grew. Separate the price effect from the volume effect in your internal analysis so you know what is really happening.

Public consumption-based companies handle this with cohort definitions based on actual revenue rather than contract run-rate. Snowflake's annual report for fiscal 2026 describes a two-year measurement period: it defines a cohort of capacity-contract customers active in the first month of the first year, then divides that cohort's product revenue in the second year by its product revenue in the first year. Customers who stopped using the platform stay in the cohort at zero. You don't need to copy that method, but it is a good model for a usage-based definition that is transparent and repeatable. My usage-based pricing model page covers how pricing choices feed into revenue forecasts.

Common calculation mistakes

These are the errors I see most often when reviewing founder metrics:

  • Including new customers. Adding revenue from customers who signed during the year turns NRR into a growth rate. It's the single most common error.
  • Measuring the wrong period. A monthly or quarterly NRR is not the same as annual NRR. If you report a shorter period, label it clearly.
  • Changing the ARR definition. If usage was counted one way last quarter and another way this quarter, the trend is meaningless.
  • Dropping churned customers from the cohort. Churned customers must stay in the denominator. Removing them inflates the result.
  • Counting one-time fees. Implementation fees, professional services and one-off credits are not recurring revenue and should be excluded.
  • Reporting only a blended number. A blended NRR can hide a weak SMB segment behind one large expanding enterprise account. Cut it by segment, cohort or plan.

A simple control helps: the ARR bridge should tie. Opening ARR plus new, expansion, contraction and churn should equal closing ARR. If the bridge doesn't tie, the NRR inputs are wrong.

How NRR connects to your other metrics

NRR is not a standalone number. It drives several others investors look at:

  • Burn multiple. Expansion reduces the cost of each net new dollar of ARR, while churn and contraction raise it. My guide to burn multiple for AI startups shows how the two link.
  • Gross margin. Expansion that comes from heavy usage can increase compute cost faster than revenue. Check margin after inference for expanding accounts, not just for the base.
  • Forecast credibility. If your plan assumes strong expansion from existing customers, the historical cohort data has to support it.

How to report NRR to investors

A clean presentation includes:

  1. Annual NRR and gross revenue retention, on a trailing twelve-month basis, with the method in a footnote.
  2. A cohort chart showing revenue by starting cohort over time.
  3. A split between price, seats and usage expansion where you can measure it.
  4. NRR by segment or plan, and the share of ARR in the largest accounts.
  5. A clear statement of how usage and pilots are treated.

Put the definition in your board pack and investor updates so the number doesn't change meaning between meetings. My investor reporting page covers how to structure those updates. If you are preparing for a raise and want a second opinion on how your retention metrics will read in diligence, you can book a call.

FAQ

Is net revenue retention the same as net dollar retention?

Yes. The two names are used interchangeably. Some companies report a dollar-based net retention rate; the calculation is the same idea.

What is a good NRR for an AI startup?

There is no single target. Above 100% means the existing base grows on its own, which most investors want to see. What counts as strong depends on segment, contract size and pricing model, so compare against companies with similar customers.

Should I calculate NRR monthly or annually?

Track it monthly internally if it helps, but report annual NRR to investors. It is the convention investors usually expect and is less affected by short-term usage swings.

Can I calculate NRR before I have a full year of customers?

You can calculate a shorter-period retention rate, but label it clearly. A twelve-month NRR needs customers who have been with you for twelve months.

Does expansion from price increases count?

Yes, price increases from existing customers count as expansion. Disclose large price changes separately so investors can see how much retention comes from usage and how much from pricing.

Work with me

Want this applied to your numbers? Book a 30-minute intro call.